| name | repo-visuals-retro |
| description | Retrospective meta-skill for the `repo-visuals` skill. Reads accumulated evaluation logs from past runs, spots patterns (recurring low-score criteria, repeated iteration failures, unsatisfied requests), consults other expert skills (skill-creator, frontend-design, etc.) where relevant, and proposes concrete edits to `repo-visuals/SKILL.md` as a reviewable diff. Runs on-demand, not per run. |
repo-visuals-retro
The repo-visuals skill scores every run on five criteria (see repo-visuals/SKILL.md §6). This meta-skill turns that accumulated evidence into skill improvements.
When to invoke
- You have ≥ 5 runs logged in
./evaluations/runs/ in the user's working directory (before that, the sample is noise).
- Something feels off across multiple runs and you want a structured look.
- A major change to
repo-visuals is planned and you want the evidence base first.
Do not run this every session. Retros are more valuable with accumulated samples.
Inputs
- The
repo-visuals skill's SKILL.md — the current skill definition (sibling skill in the same plugin)
./evaluations/index.md in the user's working directory — curated aggregate from previous retros./evaluations/runs/*.md in the user's working directory — all per-run raw evaluations since the last retro- Available expert skills in the environment (detected, not assumed)
Workflow
1. Read & tabulate
Read all inputs. Build a small table:
- Average score per criterion, overall and by repo-type bucket (CLI, library, web-app, etc. — infer from scan metadata in each run)
- Variance per criterion (high variance = the skill is inconsistent on that axis)
- Free-text feedback clustered by theme
2. Identify patterns
Name each pattern concretely:
- "Hero moment delivery averaged 2.6 / 5 across the 4 CLI-tool runs but 4.2 / 5 across library runs"
- "Technical polish dropped whenever the stage was < 400 px tall — type legibility likely"
- "3 runs had identical user feedback: 'loop seam is jarring'"
3. Consult expert skills
For each pattern, consult the relevant expert skill(s) if available in the environment. Examples:
- Visual / design critique →
frontend-design skill. Feed it the final HTML + screenshots from low-scoring runs; ask for specific design-rule violations.
- Skill structure / prompt design →
skill-creator skill. Ask whether the skill's Phase N language is ambiguous or missing a step.
- Evaluation methodology → any
evaluate-plugins style skill. Ask whether the 5-criterion scorecard is well-calibrated.
- Domain-specific knowledge → e.g. an ast-graph / codebase-compare skill if patterns suggest weaknesses in the scan phase.
Detect which skills are actually available — don't reference non-existent ones.
4. Propose edits
Write a diff-style proposal: for each pattern, which section of the repo-visuals skill's SKILL.md changes and why. Example:
Pattern: Hero moment delivery weak on CLI tools (2.6/5 avg over 4 runs).
Hypothesis: §1.4a probes don't surface CLI-specific hero moments
(install speed, clear command output, shell ergonomics).
Proposed edit to §1.4a: add a CLI-tool branch to the probing list with
3 tool-specific prompts.
Show the proposed diff. Wait for user approval before applying.
5. Apply & update aggregate
On user approval:
- Edit the
repo-visuals skill's SKILL.md with the approved changes
- Append a retro summary to
./evaluations/index.md (date, patterns identified, edits applied)
- Move the raw run files that informed this retro into
./evaluations/runs/processed/ so the next retro only sees new samples
Outputs
- Updated
repo-visuals skill SKILL.md
- Updated
./evaluations/index.md with retro summary
- Moved processed run files
What this skill does NOT do
- Does not run
repo-visuals itself
- Does not auto-apply edits without user approval
- Does not invent criteria — it only improves the skill against its existing scorecard
